Sentiment Analysis of the News Media on Artificial Intelligence Does Not Support Claims of Negative Bias Against Artificial Intelligence
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
OMICS A Journal of Integrative Biology
Volume 23, Number 0, 2019
Research Article
ª Mary Ann Liebert, Inc.
DOI: 10.1089/omi.2019.0078
Sentiment Analysis of the News Media
on Artificial Intelligence Does Not Support Claims
of Negative Bias Against Artificial Intelligence
Colin Garvey and Chandler Maskal
Downloaded by Rensselaer Polytechnic Institute NERL from www.liebertpub.com at 07/31/19. For personal use only.
Abstract
Artificial intelligence (AI) is a hot topic in digital health, as automated systems are being adopted throughout
the health care system. Because they are still flexible, emerging technologies can be shaped significantly by
media representations as well as public engagement with science. In this context, we examine the belief that
negative news media coverage of AI—and specifically, the alleged use of imagery from the movie Terminator—is
to blame for public concerns about AI. This belief is identified as a potential barrier to meaningful engagement
of AI scientists and technology developers with journalists and the broader public. We name this climate of risk
perception the ‘‘Terminator Syndrome’’—not because of its origins in the movie of the same name per se, but
because such unchecked beliefs can terminate broad public engagement on AI before they even begin. Using
both quantitative and qualitative approaches, this study examined the hypothesis that the news media coverage
of AI is negative. We conducted a sentiment analysis of news data spanning over six decades, from 1956 to
2018, using the Google Cloud Natural Language API Sentiment Analysis tool. Contrary to the alleged negative
sentiment in news media coverage of AI, we found that the available evidence does not support this claim. We
conclude with an innovation policy-relevant discussion on the current state of AI risk perceptions, and what
critical social sciences offer for responsible AI innovation in digital health, life sciences, and society.
Keywords: artificial intelligence (AI), digital health, risk governance, technology policy, AI and risk, public
engagement
Introduction present) (Garvey, 2018c; Matsuo, 2015; Ziad and Emanuel,
2016). In between these peaks, there were periods of ‘‘troughs’’
A rtificial intelligence (AI) has become a hot topic
in a wide range of science and technology fields. AI
is expected to broadly impact the digital transformation of
characterized by low optimism and lower funding, also
known as ‘‘AI Winters’’ (Fast and Horvitz, 2017; Hendler,
2008).
health care, not to mention automation in allied fields such The GOFAI paradigm was driven by symbolic logic to
as biomedical research, data science, and integrative biology. make ‘‘machines who think’’ (Haugeland, 1985; McCor-
Impacts on digital health powered by AI and automation are duck, 1979, 2004). Later, expert systems narrowed the focus
already manifesting at multiple levels in medical research from general intelligence to human expertise in specific
ranging from study design, data collection, and real-time domains, such as chemistry and medicine (Buchanan and
data analysis to societal applications of Big Data (Bohannon, Shortliffe, 1985; Feigenbaum et al., 1988). The current ML
2015a; Char et al., 2018; Forsting, 2017; Garvey, 2018d; Just paradigm extrapolates patterns directly from Big Data, usually
et al., 2017; Kaiser, 2018; Meyer et al., 2018). through a training period involving millions of trial-and-error
The current enthusiasms for AI in life sciences and health loops ( Jordan and Mitchell, 2015).
care can be better understood in the following brief historical This requires considerable computational power and
context, however. AI has evolved under three broad para- memory, which is why the recent successes of ML algorithms
digms or ‘‘peaks’’ spanning over six decades: (1) good-old- in image (e.g., in radiology) recognition and classification
fashioned AI (GOFAI) (1950–60s), (2) ‘‘expert systems’’ (Esteva et al., 2017), speech recognition and language trans-
(late 1970–80s), and (3) ‘‘machine learning (ML)’’ (2010– lation (Hirschberg and Manning, 2015), and games like Go and
Science and Technology Studies Department, Rensselaer Polytechnic Institute, Troy, New York.
12 GARVEY AND MASKAL
poker (Brown and Sandholm, 2018; Moravčı́k et al., 2017; should also be borne in mind to prevent such exercises from
Silver et al., 2016, 2017, 2018) owe as much to recent hard- being transformed into hollow pageantry to improve public
ware innovations as clever programming (Stone et al., 2016). relationships or a ‘‘tick the box’’ exercise devoid of mean-
Human values and politics (i.e., the constitution and con- ingful exercises of technological democracy (Stilgoe et al.,
testation of power) manifest themselves in various forms and 2014; Wynne, 1992).
contexts in science and technology. Scholars on the politics Allowing for dissent and disagreement is valuable to
of knowledge production have for decades described the achieve critical and reflexive perspectives at the science and
ways in which human values and power decisively shape the society interface (Garvey and Chard, 2018; Lindblom, 1990).
entire trajectory of science and emerging technologies such Moreover, forced consensus on emerging technologies does
as AI—from new idea conception to implementation science not bring about sustainable or responsible innovation (Sar-
(Agalianos, 2006; Bijker and Law, 1992; Bradshaw et al., ewitz, 2011, 2015).
2013; Campbell, 2005; Eubanks, 2017; Forsythe, 2001; Hard- Moving forward on AI and society research, there are
ing, 2006; Hoffman, 2017; Sabanović, 2014; Winner, 1980). barriers to public engagement (Garvey, 2018a). Many in the
Yet public engagement in science and technology, if de- AI technology community hold an antagonistic stance toward
signed with guidance from critical social science, can help the media. This is not unique to AI; the contentious rela-
surface the human values and power acting on emerging tionship between science and the media has been studied for
Downloaded by Rensselaer Polytechnic Institute NERL from www.liebertpub.com at 07/31/19. For personal use only.
technologies, thereby contributing to transparency and ac- decades (Nelkin, 1987, 1989, 1998). As Nelkin points out,
countability in science, as well as technological democracy scientists often ‘‘interpret critical reports about science or
(Barber, 1998; Carroll, 1971; Dotson, 2017; Eglash and technology as evidence of an anti-science or antiestablish-
Garvey, 2014; Feenberg, 2017; Sclove, 1995; Stilgoe et al., ment bias’’ and respond by attempting to control journalistic
2014; Woodhouse and Patton, 2004; Wynne, 2006). access (Nelkin, 1987: 155). For example, the Conference on
While systems science applications forge ahead in medi- Neural Information Processing Systems (NeurIPS, formerly
cine (Dzobo et al., 2018; Konstorum et al., 2018), the suite of NIPS), currently the largest AI conference with over 8000
AI technologies continues to be championed and contested attendees (Shoham et al., 2018), banned journalists from the
in parallel. In this context, several questions are necessary to tutorials and workshops in 2017, and from the workshops in
understand the AI future(s). 2018 (Shead, 2018).
The belief that negative media coverage of AI—and spe-
Will AI scientists build ‘‘AI for Social Good,’’ as many
cifically, news media’s alleged use of imagery from the
industry leaders (Dean and Fuller, 2018; Horvitz, 2017;
movie Terminator—is to blame for public concerns about AI
Nadella, 2016) and academics claim (Floridi et al.,
has been widespread in the AI community for years (Shead,
2018; Taddeo and Floridi, 2018)?
A wide range of companies, think tanks, interest groups,
2018). This belief poses a nontrivial ‘‘psychocultural bar-
rier’’ (Dotson, 2015) to broader public engagement on AI
and government agencies have taken up AI for Social
(Garvey, 2018a), thereby reducing democratic societies’ ca-
Good as an organizing theme. Why then is the public
pacities for responsibly governing the manifold risks posed
allegedly concerned about and distrustful of AI? For ex-
by AI technologies (Garvey, 2018b).
ample, one study found that the majority of Americans is
We name this climate of risk perception the ‘‘Terminator
‘‘worried’’ about AI’s effect on the job market (72%), the
Syndrome’’—not because of its origins in the movie of the
impact of hiring algorithms (67%), and the development
same name per se, but because such beliefs can terminate
of driverless cars (54%) (Pew Research Center, 2017).
broad public engagement with AI before it even begins.
The tech industry now recognizes that a lack of public
Using both quantitative and qualitative approaches, this
trust may pose one of the greatest barriers to AI adoption
study examined the hypothesis that news media coverage of
in areas from driverless cars to health care (AAA, 2018;
AI is negative. We conducted a sentiment analysis of news
Davenport, 2018; Dujmovic, 2017; Wakefield, 2018).
Finally, how and under which epistemologies will the
data spanning over six decades, from 1956 to 2018, using the
Google Cloud Natural Language API Sentiment Analysis
concerns of the public about AI be addressed? Will the
tool. We found that the available evidence does not sup-
AI community’s attempts to ameliorate public concerns
port the claim of a bias toward negative sentiment in the
and rectify the ‘‘trust crisis’’ (Agency Staff, 2018) al-
news media coverage of AI. We conclude with an innova-
low for critical approaches to technology assessment,
tion policy-relevant discussion on the current state of AI
or remain constrained to narrowly framed market effi-
risk perceptions, and what critical policy studies offer for
ciency arguments?
‘‘AI and society’’ scholarship and responsible innovation for
For answers, broad public engagement with AI is essential. emerging applications of AI in health care, life sciences, and
Such engagement exercises should be mindful of politics society.
of knowledge production, and epistemological (Özdemir and
Springer, 2018) and other forms of diversity (Callahan et al.,
Materials and Methods
2016) in the engagement practices themselves, not only be-
cause the field of AI itself is in a gender (Simonite, 2018) and The term ‘‘artificial intelligence’’ dates to 1956, when it
racial ‘‘diversity crisis’’ (Snow, 2018) but also because of the was first used in the title of a conference at Dartmouth Col-
cognitive and democratic benefits such diversity brings to lege (McCarthy et al., 2006). Because this date is generally
collective decision making (Hong and Page, 2012; Land- accepted as the origin point for the AI field (Crevier, 1993),
emore, 2013; Lindblom and Woodhouse, 1993). The politics we decided to analyze the period from 1956 to 2018.
of and the power embedded in public engagement (who is Traditional approaches to sentiment analysis of the news
represented, included, or excluded in engagement, and why?) media typically involve manually classifying articles from aSENTIMENT ANALYSIS OF THE NEWS MEDIA ON AI 3
selected corpus, delimited either by temporal range or article ternational Herald Tribune, Reuters, CNBC, International
availability (Martin, 1993). However, our intention to analyze NYT, and Internet Video Archive (NYT, 2019).
news article sentiments over AI’s entire history, combined The news aggregator News API was chosen for breadth
with the recent explosion of news articles about AI, made search, as it offers articles from over 30,000 different sources,
manual approaches to sentiment analysis low throughput and but only includes articles published within the last 3 months
thus inappropriate for this study. We therefore decided to au- of the search (for unpaid users) (News API, 2019).
tomate the task by employing a high-throughput methodology With historical depth and contemporary breadth datasets,
utilizing natural language processing (NLP) software to per- we believe the source material presented in this study to be an
form sentiment analysis on a larger (what is essentially Big adequate representation of a broad mediasphere. The NYT
Data) corpus of AI news articles than could be feasibly coded has long been regarded as the flagship news outlet in the
by hand. To the extent that heuristic search and NLP both USA, a publication of record known for relatively minimal
emerged from AI (Stone et al., 2016), this study used basic AI partisan bias and broad coverage of events. In contrast, News
tools to study the news media coverage of AI itself. This study API aggregates a diversity of news sources, from mainstream
was approved by the authors’ institutional review board (IRB). news outlets to technology blogs, thus ensuring that this study
We began by selecting news sources to establish a cor- captures media reflective of broader post-print news and In-
pus of articles. We accessed online news sources through ternet cultures.
Downloaded by Rensselaer Polytechnic Institute NERL from www.liebertpub.com at 07/31/19. For personal use only.
their application programming interfaces (APIs), a common Using the search queries ‘‘A.I.’’ and ‘‘artificial intelli-
method for aggregating news from multiple sources into gence,’’ we retrieved 8470 articles from NYT over the period
streaming feeds at a single location (e.g., Google News). Our from 1956 to March 27th, 2018, and 3906 articles from News
review of available news API services ultimately informed API over the period from December 14th, 2017, to March
the subsequent study design. To compensate for potential 14th, 2018, for a total of 12,376 articles. Figure 1 shows an
bias in coverage, we had planned to populate our corpus of overview of the data collection process.
articles from sources spanning the full ideological spectrum, The news search function examined the body text, head-
expanding out from allegedly neutral news organizations line, and byline of each article for the query text. When a
to include media outfits on both ends (e.g., conservative and match is found, the API only returns the article’s headline,
progressive) of the political spectrum. However, we found a ‘‘snippet’’ of the body text (usually the first line of the
that most news outlets had either commercialized or other- article), and metadata (such as date, section, and source).
wise restricted access to their APIs. Although the articles’ body text was unavailable, social sci-
We therefore adopted a research design utilizing depth and entific research on news coverage of science has empha-
breadth searches on separate sources of AI news media. The sized how headlines not only concisely convey the emotional
New York Times (NYT) API was chosen for depth search, as and informational content of entire articles but also play
it offers news dating from 1891 to the present, as well as a important roles in forming public opinion (Nelkin, 1987).
limited selection of articles from Associated Press, The In- Therefore, this study utilized a simplifying assumption that
FIG. 1. Flowchart of the data collection and preparation process.4 GARVEY AND MASKAL
article sentiment can be approximated using only headline human intelligence (HI) (Özdemir, 2019) with reference to
and snippet data. tacit social context (Dreyfus, 1992; Taube, 1961) in media
After removal of duplicate articles, we had 11,915 articles analysis. For example, the NYT dataset contained many ar-
in two datasets: NYT (n = 8427) and NewsAPI (n = 3488). ticles from the 1980s on ‘‘artificial insemination,’’ which
We made the assumption that if an article’s headline and described people looking for sperm donors with high ‘‘in-
snippet did not include the query terms ‘‘artificial intelli- telligence.’’ These articles were not included because they
gence’’ or ‘‘A.I.,’’ and it did not appear relevant to AI as are obviously not related to AI. In addition, the search queries
suggested by its tacit sociotechnical context (see examples returned some articles headlined by ‘‘Artificial Intelligence,’’
below), then it should be considered a false positive and re- which actually described false or ‘‘artificial’’ military intel-
moved from the dataset. ligence; these were not included.
However, ‘‘AI’’ is itself a broad concept and a fluid field The data cleaning process resulted in two final datasets
(Stone et al., 2016). This made rules for determining the (NYT = 913, NewsAPI = 2359). The sentiment of all 3272
relevance of each article challenging to define. Ultimately, articles (headline plus snippet) was then analyzed using
because this study was interested in media representations Google Cloud Natural Language Sentiment Analysis tool.
and sentiments of AI, we decided that articles containing Accordingly, for each article, Google’s Sentiment Analy-
cognate terms such as ‘‘driverless car,’’ ‘‘robot,’’ ‘‘thinking sis tool returns two numerical values that can be used ‘‘to
Downloaded by Rensselaer Polytechnic Institute NERL from www.liebertpub.com at 07/31/19. For personal use only.
machine,’’ and so on should be considered relevant. determine the overall attitude (positive or negative) ex-
For example, a headline such as ‘‘Ava of ‘Ex Machina’ Is pressed within the text’’ (Google, 2019). The first score di-
Just Sci-Fi (for Now)’’ (NYT, 5/21/2015) does not explicitly mension is a ‘‘sentiment score’’ between -1.0 (negative) and
mention ‘‘artificial intelligence’’ or ‘‘A.I.,’’ but it does refer +1.0 (positive), which ‘‘corresponds to the overall emotional
to a robot from a popular sci-fi movie about AI. Moreover, the leaning of the text.’’ We note, however, that Google offers no
article’s snippet, ‘‘A question most techies don’t seem to want explanation of how these associations are formed. Sentiment
to answer: who is making sure that all of this innovation does scores in the negative range (-) indicate the presence of text
not go drastically wrong?’’ is directly relevant to the themes of associated with overall negative emotions, and higher scores
fear, sensationalism, and danger surrounding AI that this study in the positive range (+) with positive emotions. A positive
intended to examine. Therefore, it was decided that this article score in the first dimension does not necessarily mean,
and others like it should be included in the final dataset. Ta- however, that the article contained no negative emotions in
ble 1 shows the final selection of the relevant terms. certain sections, but rather that the overall sum of the emo-
The two datasets were then cleaned of false positives, as tions was positive.
noted above, and reviewed by the authors. We started the The second score dimension is a ‘‘magnitude score’’ be-
analyses with the oldest article and moved forward toward tween 0.0 and infinity that ‘‘indicates the overall strength of
the newest, read the headline first, then the snippet, and the emotion noted (both positive and negative) with the given
looked for the query terms ‘‘A.I.,’’ ‘‘AI,’’ or ‘‘artificial in- text’’ (Google, 2019). The magnitude score is calculated as a
telligence.’’ If these terms were not found in either the sum of all sentiment, both positive and negative; in other
headline or snippet, we carried out a qualitative analysis to words, it is the absolute value of sentiment in the text.
verify and decide if the article was related to AI and attendant Therefore, if not 0.0, the magnitude score is always positive.
sentiments, using the terms in Table 1 as guidelines. In addition, it increases with each occurrence of sentiment in
The search queries used in this study returned some un- a text, so longer texts will likely have higher magnitudes,
anticipated results that attest to the importance of utilizing with no upper limit.
One limitation of Google’s Sentiment Analysis tool is that
scores of zero in the sentiment dimension can indicate either
Table 1. Terms Used to Select Relevant neutral sentiment throughout the text or the presence of both
Articles for This Study positive and negative sentiments that cancel each other out,
resulting in an overall sentiment of zero. Here, the second
Terms not magnitude score dimension can be used to distinguish be-
Terms included included tween the two possibilities. Articles with both sentiment
and magnitude scores near zero suggest truly neutral text,
Alternative terms for AI: thinking Drone whereas near-zero sentiment with high magnitude scores
machines, smart machines,
intelligent machines, etc. suggest strong contrasting sentiments within the text.
AI techniques: machine learning, Supercomputer For example, the article ‘‘Happy Birthday, HAL; What
deep learning, neural networks, Went Wrong?’’ (NYT, 1/12/92) returned a modest sentiment
expert systems, etc. score of +0.1 in the positive sentiment range and a high
Specific AI systems: IBM Watson, Artificial magnitude of 2.8. This scoring presumably reflects the
DeepMind AlphaGo, Apple’s Siri, insemination emotional contrast between ‘‘Happy’’ and ‘‘Wrong,’’ whose
Amazon Alexa, Microsoft Cortana, positive and negative sentiments canceled out in the first
etc. dimension, while the strength of emotion in both words
Applications of AI: driverless cars, ‘‘Artificial’’ (i.e., (happy and wrong) contributed to the total value of the sec-
self-driving cars, autonomous cars, false) military ond magnitude score dimension.
etc. intelligence
Cognate terms: robot (also ‘‘-bot,’’ The sentiment and magnitude score data were then vi-
‘‘chatbot,’’ ‘‘sexbot,’’ etc.) sualized using Tableau (Tableau Software, 2019), a data
visualization software package available for free under edu-
AI, artificial intelligence. cational license.SENTIMENT ANALYSIS OF THE NEWS MEDIA ON AI 5
Results and Discussion Interestingly, the average sentiment score of positive ar-
ticles was nearly identical across NYT (0.3876) and News-
News media is a pillar of technological democracy and an API (0.3857) datasets. Moreover, the average negative score
important conduit for public engagement in science. Because was also quite close in both NYT (-0.3108) and NewsAPI
news media influences both publics’ and practitioners’ per- (-0.2926) datasets.
ception, understanding, and beliefs about emerging technol- That the vast majority of articles display some sentiment,
ogies such as AI and their medical applications, the emotional whether positive or negative, suggests an appreciable degree
sentiment of news articles can shape the development and of polarization in the sentiment space of AI news coverage.
trajectory of digital health innovations. Moreover, the similarity of average positive and negative
We found that across both depth and breadth datasets, AI sentiment scores across the depth and breadth news corpuses
news coverage demonstrates robustly positive sentiment. further suggests that this polarization may be temporally
Figure 2 shows the count of articles with positive senti- robust across a broad mediasphere. One implication for
ment (>0), negative sentiment (0), negative sentiment (6 GARVEY AND MASKAL
Downloaded by Rensselaer Polytechnic Institute NERL from www.liebertpub.com at 07/31/19. For personal use only.
FIG. 3. Total count of NYT articles by year. NYT, The New York Times.
robustly positive across more than five decades, as are av- technologies like Google’s sentiment analysis tool (see our
erage sentiment magnitude scores. discussion of this problem below).
Qualitative analysis using HI through discussions among The extreme fluctuation of scores from 1965 to 1980 re-
the authors could not discern why the earliest article in the flects, in part, the paucity of articles from that period. The
dataset (‘‘Engineers Hailed for Space Work,’’ NYT, 8/23/ next lowest average sentiment (-0.031) occurs in 1988, fol-
1963), which mentioned the ‘‘Artificial Intelligentsia,’’ re- lowing the onset of ‘‘AI Winter’’ in 1987 (Fast and Horvitz,
ceived a negative sentiment score (-0.4). Anomalies such as 2017), with multiple lead stories reporting on the field’s failure
this highlight one shortcoming of using black boxed AI to live up to its promises (Garvey, 2018c). For example:
FIG. 4. Average sentiment and magnitude of NYT articles per year. Average sentiment is shown above and average
magnitude below.SENTIMENT ANALYSIS OF THE NEWS MEDIA ON AI 7
Setbacks for Artificial Intelligence: A major retrenchment is Uber of Using Stolen Technology’’ (NYT, 2/24/2017), which
occurring in the artificial intelligence industry, dashing the received a sentiment score of -0.9 and a magnitude of 0.9.
hopes of many companies that thought they would prosper by The highly publicized legal battle that followed from Goo-
providing the technology to make computers ‘‘think.’’ (NYT, gle’s accusations revealed a disturbing level of recklessness
3/4/1988)
in the driverless industry, which presumably contributed to
The lowest average annual sentiment score in the entire public perceptions of AI (Duhigg, 2018; Lee, 2018; Somer-
depth corpus (-0.067) occurs in 1996, in the middle of the AI ville and Levine, 2017).
Winter of the 1990s, before picking back up in with IBM Figure 5 displays the highest (peak), average, and lowest
Deep Blue’s 1997 win over Gary Kasparov (Campbell et al., (trough) sentiment scores, as well as average magnitude
2002; Kasparov and Greengard, 2017). However, coverage in scores, per day, for the NewsAPI breadth corpus. There were
1996 was scarce, with only three articles, and the low average more dramatic fluctuations in the trough sentiment scores
score appears to be impacted by an idiosyncrasy of Google’s than the peak scores in a given day; indeed, on 4 days (1/1, 1/
tool, which ranks words like ‘‘cockroaches’’ as extremely 13, 2/10, and 2/17), for example, there were no articles with
negative (see discussion below). negative sentiment, and on 2 days, the trough sentiment score
Taken together, Figures 3 and 4 offer an important take- was in the positive value range (1/13, 2/10). This suggests
away from this study: despite the vast increase in AI news that media coverage of AI is more consistently positive with
Downloaded by Rensselaer Polytechnic Institute NERL from www.liebertpub.com at 07/31/19. For personal use only.
coverage over the last decade, average sentiment and mag- greater stability in the positive scores, and only periodically
nitude remain positive. Moreover, the most negative articles negative.
from this period reflect real-world events rather than sensa- The average magnitude scores were also high on most days
tional speculation or fearmongering. For example, compar- and well over 0.500, suggesting most articles contain senti-
atively low average sentiment score for 2017 (0.05) is due, in mental valence (polarity), whether positive or negative.
part, to the article, ‘‘Google Self-Driving Car Unit Accuses However, the top five magnitude peaks (12/25/17, 1/6, 1/14,
FIG. 5. Daily (high, average, and low) sentiment scores and average magnitude scores for NewsAPI articles. API,
application programming interface.8 GARVEY AND MASKAL
2/4, and 2/11) occurred on days with a positive average average sentiment is positive (0.140), while the smallest cell
sentiment, which implies that spikes in emotional magnitude in the bottom right represents a source (Zohosites.com) that
might correlate with positive sentiment more strongly than only published 1 neutral article (0.0). Overall, Figure 6
with negative, at least within the period analyzed. Indeed, the qualitatively depicts a positive relationship between the
average sentiment score over the 3-month time span never number of articles published and the positive sentiment. That
dropped below zero. is, the sources that provided more AI coverage offered, on
For the NewsAPI data, the lowest average daily sentiment average, more positive coverage.
(0.0) occurred on March 3rd, 2018. Of the 10 articles published For the most part, sources with greater coverage were
that day, 5 were positive, 1 was neutral, and 4 were negative. also more recognized and reputable sites (e.g., Slashdot.org,
The two most negative articles, from Yahoo.com (-0.8) and CNBC, Forbes.com, Google News, NYT), whereas sources
Fortune (-0.3), each carried the headline, ‘‘Elon Musk Blasts with fewer articles were mostly news blogs. The exceptions
Harvard’s Steven Pinker Over Comments Dismissing the to this general qualitative trend are worth noting, however, as
Threat of Artificial Intelligence’’ (3/3/2018). Musk has been they suggest there appears to be no clear connection be-
one of the foremost figures in raising the alarm about AI tween political orientation and sentiment regarding AI.
(Anderson, 2014; Breland, 2017; Devaney, 2015). However, The largest source with a negative average sentiment score
our sampling of the news that day shows that even when a is the contested extremist media outlet Breitbart.com (12
Downloaded by Rensselaer Polytechnic Institute NERL from www.liebertpub.com at 07/31/19. For personal use only.
dystopian view of AI is reported on, its sentiment is counter- articles, average sentiment score = -0.017). Yet an equally
balanced by multiple positive articles in the mediasphere. contested extremist site, Infowars.com, was one of the larger
Figure 6 displays a treemap of the 659 different data news sources in our dataset with 38 articles, mostly com-
sources that provided the 2359 articles for the NewsAPI prised blog posts about recent market research that scored a
breadth corpus. The size of each cell corresponds to the positive average sentiment of 0.376. Identification with po-
number of articles from each source, and the shading to the litical extremism does not seem to indicate a clear negative or
averaged sentiment scores of the articles from that source. positive bias toward AI.
For example, the largest cell in the upper left represents a By contrast, Sciencemag.org, the official website of the
source (Youbrandinc.com) that published 152 articles whose journal Science and arguably a highly reputable source of
FIG. 6. Treemap of the 659 different data sources for the NewsAPI breadth corpus. Cell size corresponds to article count,
and shading to average sentiment score.SENTIMENT ANALYSIS OF THE NEWS MEDIA ON AI 9
scientific information, had five articles with a negative However, many of the lowest scoring articles were indeed
average sentiment of -0.2. Our qualitative analysis of the negative, covering a range of controversial issues highly
articles found no evidence of unfair negative bias. The sen- relevant to AI today: billion-dollar theft of corporate IP
timent score in the latter site was brought low by articles (NYT, 2/24/17); Facebook’s attempt to predict which users
about real problems facing AI at the moment, with two on are at risk of suicide (CNBC, 2/21/18); failed safety systems
the ‘‘reproducibility crisis’’ emerging in AI (‘‘Missing data in driverless cars resulting in a pedestrian’s death (NYT,
hinder replication of artificial intelligence studies,’’ 2/15/ 3/21/18); the poor performance of AI-powered hedge funds
2018, -0.3 and ‘‘Artificial intelligence faces reproducibility (Investopedia.com, 3/14/18); issues of consent raised by sex
crisis,’’ 2/15/2018, -0.2) and one on the authoritarian polit- robots (NYT, 7/17/17); sexual misconduct by AI executives
ical applications of AI in China (‘‘China’s massive invest- (The Hill, 12/16/17); the malicious use of AI (Reason.com,
ment in artificial intelligence has an insidious downside,’’ 2/21/18); the threat of authoritarian AI technologies in China
2/8/2018, -0.8). (Metafilter.com, 12/15/17); and so on.
Although the Google NLP sentiment analysis tool is ar-
Contextualizing the findings guably a simple one, it is not immediately clear that more
Our most salient finding that AI news coverage, by and complex tools are necessarily appropriate to answering the
large, is robustly positive casts doubt on the belief (Guizzo questions posed by our study about the sentiment of media
Downloaded by Rensselaer Polytechnic Institute NERL from www.liebertpub.com at 07/31/19. For personal use only.
and Ackerman, 2016; Scharre, 2017; Shead, 2018; Williams, coverage of AI. For example, Fast and Horvitz (2017) con-
2015) that negative public perceptions of AI can be explained ducted a far more costly and comprehensive analysis of
by negative media coverage of AI. There are, however, ca- AI coverage in the NYT over the period 1986–2016. Using
veats. crowdsourcing and ML classifiers, they attempted to measure
First, we acknowledge that quantitative scoring of emo- ‘‘optimism versus pessimism’’ and ‘‘engagement,’’ as well as
tional ‘‘sentiment’’ is a reductive construction that fails to a number of hopes for and concerns about AI. However, the
capture contextual nuance. One could argue the score says relevant results of this sophisticated study confirm our own:
more about the programmers who created the tool than the ‘‘In general, AI has had consistently more optimistic than
text it processes, making this entire exercise meaningless. pessimistic coverage over time, roughly two to three times
However, sentiment analysis tools, including Google’s, are more over the 30-year period’’ (Fast and Horvitz, 2017).
used widely online and in industry for market forecasting and Moreover, in their sophisticated technical analysis, Fast
other forms of analysis. For example, the AI Index 2017 and and Horvitz collapse the social dimension of their data—
2018 included a sentiment analysis of AI media coverage paragraphs of text from the NYT—to stand in for public belief
(Shoham et al., 2017, 2018). Shoham et al. (2017) found itself. For example, they interpret an increase in one classi-
positive AI articles outnumbered negative during 2013–17, fier, ‘‘ethical concerns about AI,’’ to ‘‘suggest an increase
although most articles were neutral (2017: 25). in public belief that we may soon be capable of building
Shoham et al. (2018) extended those results, adding that dangerous AI systems’’ (Fast and Horvitz, 2017: 967). This
‘‘AI articles have become less neutral and more positive, study, by contrast, acknowledges the possibility of difference
particularly since early 2016, when articles went from 12% between media representations and alleged public beliefs.
positive in January 2016 to 30% positive in July 2016,’’ Indeed, that appears to be the case with AI: despite robustly
where it has remained since. However, the AI Index results positive media coverage, concerns about the technology are
do not bear a strong burden of proof because the provenance widespread among the public.
of the data used is unclear. It was obtained from a business Third, it is certainly possible that studies on larger datasets
analytics firm (Trendkite.com), but beyond that, no infor- could potentially contest our findings. Future studies should
mation is provided regarding the number of articles analyzed, analyze sentiment from many more news sources, perhaps
their sources, or the methods used. This study thus improved controlling for ideology by surveying across the political
on the AI Index study by conducting a sentiment analysis of spectrum. Ethnographic research is needed to better explore
the AI news media coverage over a longer period, while the process by which ordinary people make sense of prom-
providing transparency on data, sources, and methodology. issory technologies like AI, and the role that news media
Second, the accuracy of Google’s sentiment analysis tool plays in that understanding. However we leave that to future
can certainly be questioned. We acknowledge that other studies on the subject matter.
sentiment analysis tools might provide more nuanced scores. Having addressed those points, we now ask the follow-
The Google tool itself is a black box that does not explain ing: If most AI news is positive, why then is a majority of
how sentiments are weighted, although our examination of the public allegedly worried about and distrustful of AI? We
the most negative articles revealed some of its idiosyncrasies. consider three hypotheses.
For example, consider the following article that received one We agree it is possible that despite the mostly positive
of the lowest sentiment scores (-0.9): news coverage, a small number of highly sensational, ‘‘fear-
mongering’’ articles could have a disproportionate effect on
In Kingdom of Cockroaches, Leaders Are Made, Not Born— public perceptions of AI. Cognitive or other biases might cause
An international team of scientists has created artificial extremely negative stories (e.g., Strange, 2015) to outweigh
roaches to study ‘‘collective intelligence.’’ (NYT, 12/7/2004).
the preponderance of less remarkable positive stories in the
In this case, the low score is almost certainly an artifact of publics’ judgment. However, plausible, the burden of proof
the Google’s NLP tool, which most likely assigns very low lies with those who would advance such a hypothesis.
values to the word ‘‘cockroaches.’’ It seems unlikely that Moreover, the principle of symmetry (Bloor, 1991) urges
this article contributed significantly to negative public per- caution in invoking bias to explain public perceptions with-
ceptions of AI. out considering its role in experts’ own perceptions of AI10 GARVEY AND MASKAL
news. Biases are an inescapable fact of human cognition izing and extreme videos (Cook, 2018; Lewis, 2018; Nicas,
(Gigerenzer and Goldstein, 1996; Kahneman, 2011); hence, 2018; Tufekci, 2018)? Should private initiatives to put screens
similar explanations can be applied to both experts and lay in public schools (Kardaras, 2016; Toyama, 2015) not be
people. That is, by the same logic, one could argue experts’ reconsidered in light of the fact that Silicon Valley elites are
biases inure them to the steady stream of positive AI news, increasingly restricting their own children’s use of handheld
while making them exceptionally sensitive to the occasional digital electronics (Bowles, 2018)?
negative article, which would become overrepresented in At a minimum, we suggest that ‘‘digital health’’ should
their mental model of ‘‘AI news’’ and create the perception of be construed broadly enough to consider these and related
a biased media. questions about the role of AI technologies in human life.
A more parsimonious explanation for negative public Indeed, it seems to us that digital health cannot be understood
perceptions of AI is that the technology actually does pose outside the context of AI technologies, with their potential
considerable risks to many people, and that the public is benefits and manifold risks. This context suggests an orienta-
rightly concerned (Didier et al., 2015). Consider how after tion for the digital health research agenda. If AI poses signif-
early denials (Williams, 2015), some major tech leaders are icant risks to a nontrivial fraction of humanity (Garvey,
now conceding that AI does pose risks, such as Google CEO 2018b), then we suggest that future research in digital health
Sundar Pichai, who recently acknowledged that fears about could make the greatest impact by adopting a perspective
Downloaded by Rensselaer Polytechnic Institute NERL from www.liebertpub.com at 07/31/19. For personal use only.
AI are ‘‘very legitimate’’ (Romm et al., 2018). of ‘‘thoughtful partisanship’’ (Woodhouse et al., 2002) and
Leading AI experts are warning about the military risks explicitly focusing on the needs of those groups at greatest risk.
(Bohannon, 2015b; FLI, 2015). Senior politicians and gov- Because those at-risk populations are often also those least
ernment officials argue that AI threatens political democracy able to influence political decision making and technological
(Graham, 2017; Nemitz, 2018). Reckless corporate behavior R&D (Lindblom and Woodhouse, 1993), research oriented
has already led to the first pedestrian death by driverless car around their concerns will help to counteract the well-
(Lee, 2018; Wakabayashi, 2018). ML has been shown to documented tendency for new technologies, especially in
systematically reproduce biases in its training data (Caliskan medicine, to disproportionately benefit the wealthy and am-
et al., 2017), and such algorithmic biases have entrenched plify social inequality (Sarewitz and Woodhouse, 2003;
discriminatory practices in range of social settings, from Woodhouse and Sarewitz, 2007). We recommend science
criminal sentencing (Angwin et al., 2016; Matacic, 2018) to and technology policies for promoting ‘‘digital health’’ to be
facial recognition (Buolamwini and Gebru, 2018). In- reoriented accordingly.
vestigative journalists have shown how AI allows advertisers For digital health to be more than the next bump in the
to discriminate by race, gender, and other categories (Angwin hype cycle, a systems perspective that includes critical social
and Parris Jr., 2016; Angwin et al., 2017). science and public engagement as inputs will be crucial. AI
Finally, business people, technical experts, financial in- is increasingly a part of ordinary life, and the public is in-
stitutions, think tanks, governments, and much of the public creasingly familiar with it. Explaining away public concerns
agree that AI puts jobs at risk (Acemoglu and Restrepo, 2017; about AI as mere fear of the Terminator, stoked by an un-
Arntz et al., 2016; Brynjolfsson and Mitchell, 2017; Dutton scrupulous or sensationalist media, poses a barrier to greater
et al., 2018; Frey and Osborne, 2013; Ma et al., 2015; Pew public engagement with AI by reproducing the ‘‘deficit mod-
Research Center, 2014; World Economic Forum, 2018). el’’ of the public’s understanding of science (Bauer et al.,
What about machine takeovers and killer robots? It does 2007; Fujigaki, 2009; Ziman, 1991), which too often assumes
not appear to be the public’s primary concern. One survey ‘‘lay publics are only capable of taking sentimental, emo-
of 2000 American adults found that their ‘‘top fear or con- tional, and intellectually vacuous positions’’ (Wynne, 2001).
cern about possible A.I. threats or risks’’ was ‘‘A.I. taking Humane futures in digital health and the responsible ad-
jobs’’ (30%), rather than ‘‘A.I. turning against us’’ (8.8%) or vancement of AI technologies require that the concerns of
‘‘AI taking control’’ (6.8%) (SYZYGY, 2017). Indeed, even an informed public be heard, addressed, and incorporated into
James Cameron, director of the Terminator movies, sug- research, design, deployment, and operation. For that to
gests shifting attention to more quotidian instantiations of happen, open, transparent, and deliberative interaction be-
AI technology and their social impacts: tween AI innovators, news media, government officials, civic
People ask me: ‘‘Will the machines ever win against hu-
leaders, and diverse publics is essential.
manity?’’ I say: ‘‘Look around in any airport or restaurant
and see how many people are on their phones. The machines Conclusions
have already won.’’ (Belloni and Kit, 2017)
Using the sentiment scores calculated by the Google Cloud
As we look around, it is worth asking what ‘‘digital health’’ NLP tool, we found that the majority of AI news coverage
means—and ought to mean—in this burgeoning Age of AI, from the sources we examined is positive. If this tool is re-
where a certain class of machines has ‘‘already won.’’ What liable and accurate, and if our news sources are representa-
role should AI play in childhood development as screens tive, then our finding that a majority of AI news coverage is
become fixtures in peoples’ lives at ever earlier ages, espe- positive refutes the hypothesis that most media coverage of
cially as it becomes clear that screen time correlates with AI is negative. This casts doubt on the validity of the be-
adverse effects on young people (Hunt et al., 2018; Turel, lief, which we have called the ‘‘Terminator Syndrome,’’ that
2019; Twenge et al., 2018, 2019)? How ought guidelines for negative media coverage is largely to blame for negative
digital entertainment be set, when even a site as seemingly public perceptions of AI. Although our results alone are un-
innocuous as YouTube is now known to use AI to maximize likely to terminate the Terminator Syndrome, we hope they
screen time by tantalizing viewers with increasingly polar- contribute to facilitating greater public engagement with AI.SENTIMENT ANALYSIS OF THE NEWS MEDIA ON AI 11
Acknowledgments Bradshaw JM, Hoffman RR, Woods DD, et al. (2013). The
seven deadly myths of ‘autonomous systems.’ IEEE Intell
No funding was received in support of this research article.
Syst 28, 54–61.
Views expressed are authors’ personal opinions only and do Breland A. (2017). Elon Musk: We need to regulate AI before
not necessarily reflect those of the affiliated institutions. ‘‘it’s too late.’’ TheHill. https://thehill.com/policy/technology/
342345-elon-musk-we-need-to-regulate-ai-before-its-too-late.
Author Disclosure Statement Brown N, and Sandholm T. (2018). Superhuman AI for heads-
up no-limit poker: Libratus beats top professionals. Science
The authors declare that no competing financial interests 359, 418–424.
exist. Brynjolfsson E, and Mitchell T. (2017). What can machine
learning do? Workforce implications. Science 358, 1530–
1534.
References
Buchanan BG, and Shortliffe EH. (1985). Rule-Based Expert
AAA. (2018). AAA: American trust in autonomous vehicles Systems: The MYCIN Experiments of the Stanford Heuristic
slips. https://newsroom.aaa.com/2018/05/aaa-american-trust- Programming Project. Reading, MA: Addison-Wesley.
autonomous-vehicles-slips/. Buolamwini J, and Gebru T. (2018). Gender shades: Intersec-
Acemoglu D, and Restrepo P. (2017). Robots and jobs: Evi- tional accuracy disparities in commercial gender classifica-
Downloaded by Rensselaer Polytechnic Institute NERL from www.liebertpub.com at 07/31/19. For personal use only.
dence from US labor markets. Working paper no. 23285. tion. Proc Mach Learn Res 81, 1–15.
DOI: 10.3386/w23285. Caliskan A, Bryson JJ, and Narayanan A. (2017). Semantics
Agalianos A. (2006). The social shaping of logo. Soc Stud Sci derived automatically from language corpora contain human-
36, 241–267. like biases. Science 356, 183–186.
Agency Staff. (2018). IBM joins Apple in blasting Silicon Callahan BR, Hathaway C, and Krishnamoorthy M. (2016).
Valley for ‘‘trust crisis.’’ TechCentral. https://techcentral.co.za/ Quantitative metrics for generative justice: Graphing the value
ibm-joins-apple-in-blasting-silicon-valley-for-trust-crisis/85495/. of diversity. Teknokultura 13, 567–586.
Anderson L. (2014). Elon Musk’s artificial intelligence fear: Campbell ND. (2005). Suspect technologies: Scrutinizing the
Machines could wipe us out. Vanity Fair. www.vanityfair intersection of science, technology, and policy. Sci Technol
.com/news/tech/2014/10/elon-musk-artificial-intelligence-fear Human Values 30, 374–402.
Angwin J, Larson J, Mattu S, et al. (2016). Machine bias: Campbell M, Hoane AJ, and Hsu F. (2002). Deep blue. Artif
There’s software used across the country to predict future Intell 134, 57–83.
criminals. And it’s biased against blacks. ProPublica. https:// Carroll JD. (1971). Participatory technology. Science 171, 647–653.
www.propublica.org/article/machine-bias-risk-assessments- Char DS, Shah NH, and Magnus D. (2018). Implementing
in-criminal-sentencing machine learning in health care—Addressing ethical chal-
Angwin J, and Parris T. (2016). Facebook lets advertisers ex- lenges. N Engl J Med 378, 981–983.
clude users by race. ProPublica. https://www.propublica.org/ Collingridge D. (1980). The Social Control of Technology. New
article/facebook-lets-advertisers-exclude-users-by-race York: St. Martin’s Press.
Angwin J, Tobin A, and Varner M. (2017). Facebook (still) Cook J. (2018). The YouTube Kids app has been suggesting
letting housing advertisers exclude users by race. ProPublica. a load of conspiracy videos to children. Business Insider.
https://www.propublica.org/article/facebook-advertising- www.businessinsider.com/youtube-suggested-conspiracy-
discrimination-housing-race-sex-national-origin videos-to-children-using-its-kids-app-2018-3
Arntz M, Gregory T, and Zierahn U. (2016). The risk of auto- Crevier D. (1993). AI: The Tumultuous History of the Search for
mation for jobs in OECD countries: A comparative analysis. Artificial Intelligence. New York, NY: Basic Books.
OECD social, employment and migration working papers no. Davenport TH. (2018). Can we solve AI’s ‘trust problem’? MIT
189. DOI: 10.1787/5jlz9h56dvq7-en. Sloan Management Review. https://sloanreview.mit.edu/
Barber BR. (1998). Three scenarios for the future of technology article/can-we-solve-ais-trust-problem/.
and strong democracy. Polit Sci Quart 113, 573–589. Dean J, and Fuller J. (2018). AI for social good. Google. https://
Bauer MW, Allum N, and Miller S. (2007). What can we learn www.blog.google/outreach-initiatives/google-org/ai-social-
from 25 years of PUS survey research? Liberating and ex- good/
panding the agenda. Public Underst Sci 16, 79–95. Devaney J. (2015). Hawking, Musk sign letter warning of ar-
Belloni M, and Kit B. (2017). James Cameron sounds the alarm tificial intelligence ‘‘arms race.’’ Newsmax. www.newsmax
on artificial intelligence and unveils a ‘‘Terminator’’ for the .com/Newsfront/Stephen-Hawking-Elon-Musk-letter-artificial-
21st century. The Hollywood Reporter. www.hollywoodreporter intelligence/2015/07/27/id/659161/
.com/features/james-cameron-sounds-alarm-artificial-intelligence- Didier C, Duan W, Dupuy JP, et al. (2015). Acknowledging
unveils-a-terminator-21st-century-1043027 AI’s dark side. Science 349, 1064.
Bijker WE, and Law J. (1992). Shaping Technology/Building Dotson T. (2015). Technological determinism and permission-
Society: Studies in Sociotechnical Change. Cambridge, MA: less innovation as technocratic governing mentalities: Psy-
MIT Press. chocultural barriers to the democratization of technology.
Bloor D. (1991). Knowledge and Social Imagery. Chicago: Engag Sci Technol Soc 1, 98–120.
University of Chicago Press. Dotson T. (2017). Technically Together: Reconstructing Com-
Bohannon J. (2015a). The synthetic therapist. Science 349, munity in a Networked World. Cambridge: MIT Press.
250–251. Dreyfus HL. (1992). What Computers Still Can’t Do. New
Bohannon J. (2015b). Fears of an AI pioneer. Science 349, 252. York: MIT Press.
Bowles N. (2018). Silicon valley nannies are phone police for Duhigg C. (2018). Did Uber steal Google’s intellectual prop-
kids. The New York Times. https://www.nytimes.com/2018/ erty? The New Yorker. https://www.newyorker.com/magazine/
10/26/style/silicon-valley-nannies.html 2018/10/22/did-uber-steal-googles-intellectual-property12 GARVEY AND MASKAL
Dujmovic J. (2017). What’s holding back artificial intelligence? Botelho A, eds. Amherst: University of Massachusetts Press,
Americans don’t trust it. MarketWatch. www.marketwatch 1–236.
.com/story/whats-holding-back-artificial-intelligence-americans- Gigerenzer G, and Goldstein DG. (1996). Reasoning the fast
dont-trust-it-2017-03-30 and frugal way: Models of bounded rationality. Psychol Rev
Dutton T, Barron B, and Boskovic G. (2018). Report. Building 103, 650–669.
and AI world: Report on national and regional AI strategies. Google. (2019). Natural language API basics j cloud natural
CIFAR. Toronto, Ontario, Canada. language API. Google Cloud. https://cloud.google.com/natural-
Dzobo K, Rowe A, Senthebane DA, et al. (2018). Three- language/docs/basics.
dimensional organoids in cancer research: The search for the Graham L. (2017). Opening statement at hearing on Russian
holy grail of preclinical cancer modeling. OMICS 22, 733–748. disinformation, extremist content online. United States Se-
Eglash R, and Garvey C. (2014). Basins of attraction for gen- nator Lindsey Graham. https://www.lgraham.senate.gov/public/
erative justice. In: Chaos Theory in Politics. Banerjee S, index.cfm/2017/10/graham-opening-statement-at-hearing-on-
Erçetin SxSx, and Tekin A, eds. Dordrecht: Springer Nether- russian-disinformation-extremist-content-online
lands, 75–88. Guizzo E, and Ackerman E. (2016). When robots decide to kill.
Esteva A, Kuprel B, Novoa RA, et al. (2017). Dermatologist- IEEE Spectr 53, 38–43.
level classification of skin cancer with deep neural networks. Harding SG. (2006). Science and Social Inequality: Feminist
Nature 542, 115–118. and Postcolonial Issues. Urbana: University of Illinois Press.
Downloaded by Rensselaer Polytechnic Institute NERL from www.liebertpub.com at 07/31/19. For personal use only.
Eubanks V. (2017). Automating Inequality: How High-Tech Haugeland J. (1985). Artificial Intelligence: The Very Idea.
Tools Profile, Police, and Punish the Poor. New York: St. Cambridge, MA: MIT Press.
Martin’s Press. Hendler J. (2008). Avoiding another AI winter. IEEE Intell Syst
Fast E, and Horvitz E. (2017). Long-term trends in the pub- 23, 2–4.
lic perception of artificial intelligence. In: Proceedings of Hirschberg J, and Manning CD. (2015). Advances in natural
the Thirty-First AAAI Conference on Artificial Intelligence language processing. Science 349, 261–266.
(AAAI-17). Palo Alto, CA: AAAI Press, 963–969. Hoffman, SG. (2017) Managing ambiguities at the edge of
Feenberg A. (2017). Technosystem: The Social Life of Reason. knowledge: Research strategy and artificial intelligence labs
Cambridge, MA: Harvard University Press. in an era of academic capitalism. Sci Technol Human Values
Feigenbaum EA, McCorduck P, and Nii P. (1988). The Rise of 42, 703–740.
the Expert Company: How Visionary Companies Are Using Hong L, and Page SE. (2012). Some microfoundations of col-
Artificial Intelligence to Achieve Higher Productivity and lective wisdom. In: Collective Wisdom: Principles and Me-
Profits. New York: Times Books. chanisms. Landemore H, and Elster J, eds. New York:
FLI. (2015). Open letter on autonomous weapons. Future of Cambridge University Press, 56–71.
Life Institute. http://futureoflife.org/open-letter-autonomous- Horvitz E. (2017). AI, people, and society. Science 357, 7.
weapons/. Hunt MG, Marx R, Lipson C, et al. (2018). No more Fomo:
Floridi L, Cowls J, Beltrametti M, et al. (2018). AI4People—An Limiting social media decreases loneliness and depression. J
ethical framework for a good AI society: Opportunities, risks, Soc Clin Psychol 37, 751–768.
principles, and recommendations. Minds Mach 28, 689–707. Jordan MI, and Mitchell TM. (2015). Machine learning: Trends,
Forsting M. (2017). Machine learning will change medicine. perspectives, and prospects. Science 349, 255–260.
J Nucl Med 58, 357–358. Just MA, Pan L, Cherkassky VL, et al. (2017). Machine
Forsythe DE. (2001). Studying Those Who Study Us: An An- learning of neural representations of suicide and emotion
thropologist in the World of Artificial Intelligence. Stanford, concepts identifies suicidal youth. Nat Human Behav 1, 911.
California: Stanford University Press. Kahneman D. (2011). Thinking, Fast and Slow. New York:
Frey CB, and Osborne MA. (2013). Report. The future of em- Farrar, Straus and Giroux.
ployment: How susceptible are jobs to computerisation. Kaiser J. (2018). Cancer institute head touts big data and basic
Oxford Martin School of Business. Oxford, UK. research. Science 359, 15.
Fujigaki Y. (2009). STS in Japan and East Asia: Governance of Kardaras DN. (2016). Screens in schools are a $60 billion hoax.
science and technology and public engagement. East Asian Time. http://time.com/4474496/screens-schools-hoax/.
Sci Technol Soc Int J 3, 511–518. Kasparov GK, and Greengard M. (2017). Deep Thinking: Where
Garvey C. (2018a). A framework for evaluating barriers to the Machine Intelligence Ends and Human Creativity Begins. New
democratization of artificial intelligence. In: Proceedings of York: PublicAffairs.
the Thirty-Second AAAI Conference on Artificial Intelligence Konstorum A, Lynch ML, Torti SV, et al. (2018). A systems
(AAAI-18). Palo Alto, CA: AAAI Press, 8079–8080. biology approach to understanding the pathophysiology of
Garvey C. (2018b). AI risk mitigation through democratic high-grade serous ovarian cancer: Focus on iron and fatty
governance: Introducing the 7-dimensional AI risk horizon. acid metabolism. OMICS 22, 502–513.
In: Proceedings of the 2018 AAAI/ACM Conference on AI, Landemore H. (2013). Democratic Reason: Politics, Collective
Ethics, and Society. New York: ACM, 366–367. Intelligence, and the Rule of the Many. Princeton: Princeton
Garvey C. (2018c). Broken promises and empty threats: University Press.
The evolution of AI in the USA, 1956–1996. Technology’s Lee TB. (2018). The hype around driverless cars came crashing
Stories. www.technologystories.org/ai-evolution/ down in 2018. Ars Technica. https://arstechnica.com/cars/
Garvey C. (2018d). Interview with Colin Garvey, Rensselaer 2018/12/uber-tesla-and-waymo-all-struggled-with-self-driving-
Polytechnic Institute. Artificial intelligence and systems med- in-2018/.
icine convergence. OMICS 22, 130–132. Lewis P. (2018). ‘‘Fiction is outperforming reality’’: How
Garvey C, and Chard DS. (2018). Disrupting the AAA$. YouTube’s algorithm distorts truth. The Guardian. https://
In: Science for the People: Documents from America’s www.theguardian.com/technology/2018/feb/02/how-youtubes-
Movement of Radical Scientists. Schmalzer S, Chard DS, and algorithm-distorts-truthYou can also read
